A generous $100,000 compute infrastructure award from Fetch.AI is set to accelerate two cutting-edge research projects in the UCLA Computer Science Department, driving advances in synthetic data generation for software testing and memory-efficient large language models (LLMs).
The first project, led by Professor Miryung Kim and her team in the Software Engineering and Analysis Lab, tackles a long-standing challenge in automated software testing: the generation of synthetic data that is not realistic and fails to reflect real-world complexities. These so-called “hallucinations” in test data can limit the effectiveness of testing in data-intensive scalable computing (DISC) frameworks. To address this, Professor Kim’s team is pioneering the use of probabilistic grammar-based constrained decoding models, an innovative approach that aims to generate more realistic synthetic inputs capable of uncovering deeper bugs in complex analytics workloads.
This work builds on Professor Kim’s prior breakthroughs with tools like NaturalFuzz and NaturalSym, which introduced methods for injecting naturalness into fuzz testing and symbolic execution, respectively. NaturalFuzz uses provenance-based data mixing to create more lifelike test cases, while NaturalSym embeds constraints that mimic real-world data behaviors. With Fetch.AI’s high-performance GPU infrastructure, the team is now scaling their efforts by training probabilistic circuits and grammar-aware decoding strategies in collaboration with Professor Guy van den Broeck’s group. Their shared goal is to push synthetic testing closer to the realism of real-world data and strengthen the reliability and reproducibility of data-driven systems, which is especially crucial for modern science and engineering applications where repeatable experiments are vital.
At the same time, Professor Quanquan Gu is spearheading a second project intended to solve one of the most pressing limitations in LLMs: inference-time memory consumption. His team is investigating novel architectures based on tensor-product attention mechanisms and group-theoretic, data-dependent positional embeddings. These techniques are designed to reduce the memory footprint associated with traditional key-value (KV) caching, which is a major bottleneck in standard multi-head attention (MHA) systems used in transformers. By applying tensor decompositions, the team aspires to shrink KV cache sizes by up to a factor of 10 without sacrificing performance on key language modeling benchmarks.
Fetch.AI’s compute support allows Professor Gu’s team to rapidly prototype, train, and evaluate these new models at scale. This will ensure they are suitable for real-world deployments such as long-context language understanding, where efficient memory use is critical. Together, these two UCLA-led projects illustrate the transformative potential of academic-industry collaboration, and how advanced compute resources can unlock new breakthroughs in AI research and software engineering.